…Smith, and other providers (fixes #5763) (#5764)

* fix(analytics): capture token usage and model for Langfuse, LangSmith, and other providers

What changed
------------
- handler.ts: Extended onLLMEnd() to accept string | structured output. When
  structured output is passed, we now extract content, usageMetadata (input/
  output/total tokens), and responseMetadata (model name) and forward them
  to all analytics providers. Added usage/model to Langfuse generation.end(),
  LangSmith llm_output, and token attributes for Lunary, LangWatch, Arize,
  Phoenix, and Opik. Call langfuse.flushAsync() after generation.end() so
  updates are sent before the request completes.
- LLM.ts: Pass full output object from prepareOutputObject() to onLLMEnd
  instead of finalResponse string, so usage and model are available.
- Agent.ts: Same as LLM.ts — pass output object to onLLMEnd.
- ConditionAgent.ts: Build analyticsOutput with content, usageMetadata, and
  responseMetadata from the LLM response and pass to onLLMEnd.
- handler.test.ts: Added unit tests for the extraction logic (string vs
  object, token field normalization, model name sources, missing fields).
  OpenAIAssistant.ts call sites unchanged (Assistants API; no usage data).

Why
---
Fixes #5763. Analytics (Langfuse, LangSmith, etc.) were only receiving
plain text from onLLMEnd; usage_metadata and response_metadata from
AIMessage were dropped, so token counts and model names were missing in
dashboards and cost tracking.

Testing
-------
- pnpm build succeeds with no TypeScript errors.
- Manual: Flowise started, Agentflow with ChatOpenAI run; LangSmith and
  Langfuse both show token usage and model on the LLM generation.
- Backward compatible: call sites that pass a string (e.g. OpenAIAssistant)
  still work; onLLMEnd treats string as content-only.

Co-authored-by: Cursor <cursoragent@cursor.com>

* refactor(analytics): address PR review feedback for token usage handling

- LangSmith: Only include token_usage properties that have defined values
  to avoid passing undefined to the API
- Extract common OpenTelemetry span logic into _endOtelSpan helper method
  used by arize, phoenix, and opik providers

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(analytics): LangSmith cost tracking and flow name in traces

- LangSmith: set usage_metadata and ls_model_name/ls_provider on run extra.metadata
  so LangSmith can compute costs from token counts (compatible with langsmith 0.1.6
  which has no end(metadata) param). Infer ls_provider from model name.
- buildAgentflow: use chatflow.name as analytics trace/run name instead of
  hardcoded 'Agentflow' so LangSmith and Langfuse show the Flowise flow name.

Co-authored-by: Cursor <cursoragent@cursor.com>

* update handlers to include model and provider for analytics

* fix: normalize provider names in analytics handler to include 'amazon_bedrock'

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Henry <hzj94@hotmail.com>